Duration: 1:10:42

192 views

1

In recent years discriminative learning techniques have seen a surge of interest in the NLP community due their ability to tractably incorporate millions of dependent and linguistically rich features. In many fields, most notably information extraction, discriminative models have become the standard. In this talk I will describe a generalization of the multi-class online large-margin algorithms of Crammer and Singer (2003) to structured outputs. I apply this learning framework to the problem of extracting dependency tree representations of sentences in conjunction with a spanning tree (maximum branching) parsing framework that leads to efficient algorithms for projective and non-projective structures. I show that parsers trained under this framework can achieve state-of-the-art accuracies when combined with a rich feature set. Further more I will describe experiments displaying that these parsers are naturally extendable and can be adapted to new domains through additional features defined from information from in and out-of-domain classifiers.